How to blur license plates in equirectangular panoramas
License plates are small, high-contrast, and personal data in most jurisdictions. In a flat photo they are easy to spot; in a 360° equirectangular panorama they are one of the harder things to catch reliably. This guide explains why, and how to anonymise them without degrading the rest of your image.
Why plates slip through
A plate typically occupies a tiny fraction of a multi-thousand-pixel panorama, often at an angle, partly in shadow, and sometimes split across the left/right seam of the equirectangular projection. Detectors that run once over the whole flattened image at a fixed scale routinely miss plates that are small, rotated, or near the edges — and a single missed plate can be enough to identify a vehicle and its owner.
How multi-scale AI detection solves it
Privacy Keeper tackles plates the same way it handles faces: it samples each panorama in overlapping patches at more than one field of view, runs AI-powered license plate detection on each patch, then projects every hit back onto the sphere and merges overlapping results. Working at multiple scales means both a close-up plate and a distant one are detected, and projection-aware merging means plates near the seam are not double-counted or dropped.
Blur without wrecking the image
Anonymisation should remove the plate, not the usefulness of the panorama. A Gaussian blur is applied only to the detected region, so the surrounding street scene stays sharp. Because the original file is never overwritten and EXIF/GPS metadata is preserved, the anonymised output still drops onto your map exactly where it was shot.
Catch the ones AI can't
No detector is perfect on every frame. Two features close the gap:
- Manual review — inspect each image at full resolution with the built-in Review tool to add a blur region over any plate the model missed, or remove a false positive.
- Mask Editor — set permanent masks for areas such as the survey vehicle's own bonnet and roof, this way recurring objects and even mirrored numberplates and faces are covered automatically.
Run the whole survey at once
Batch processing with checkpoint/resume handles thousands of panoramas unattended, and each run produces an audit log you can keep as proof of anonymisation. On a 12K Mosaic 51 panorama, detection runs in about 9 seconds on an RTX 3090, with a CPU fallback when no GPU is available.